# Using mixture cure models to address algorithmic bias in diagnostic timing: autism as a test case

## TL;DR

This paper introduces mixture cure models to reduce bias in predicting autism diagnoses by accounting for differences in diagnostic timing and racial disparities.

## Contribution

The novel use of mixture cure models addresses algorithmic bias in clinical prediction by integrating time-to-event and classification frameworks.

## Key findings

- Traditional models show increased bias with wider diagnostic timing differences, while mixture cure models remain unbiased.
- Mixture cure models effectively adjust for racial disparities in autism diagnosis rates in real-world Medicaid data.
- The approach shifts focus from when a diagnosis occurs to whether it will occur, improving fairness in predictions.

## Abstract

To address algorithmic bias in clinical prediction models related to the timing of diagnosis, we evaluated the efficacy of mixture cure models that integrate time-to-event and binary classification frameworks to predict diagnoses.

We conducted a simulation and analyzed real-world North Carolina Medicaid data for children born in 2014, followed until 2023. The study evaluated traditional time-to-event and classification models against mixture cure models under scenarios with varied diagnostic timing and censoring.

Simulation results demonstrated that traditional models exhibit increased bias as diagnosis timing differences widened, whereas mixture cure models yielded unbiased estimates across varying censoring times. In real-world analyses, significant racial and ethnic variations in autism diagnosis rates were observed, with non-Hispanic White children having higher diagnosis rates compared to other groups. The mixture cure model effectively adjusted for these disparities, providing fairer and more accurate diagnostic predictions across varying levels of censoring.

Mixture cure models effectively address algorithmic bias by providing unbiased estimates regardless of variations in diagnostic timing and censoring, making them particularly suitable for conditions like autism where not all individuals will receive a diagnosis. This approach shifts focus from when an event will occur to whether it will occur, aligning more closely with clinical needs in early detection of pediatric developmental conditions.

Mixture cure models offer a promising tool to enhance accuracy and fairness in predictive modeling, especially when the outcome of interest is not uniformly observed across groups.